Re: [Starpu-devel] Problem with performance model calibration

[Mirko Myllykoski <mirkom@cs.umu.se>]

Hi,

On 2020-05-20 17:39, Samuel Thibault wrote:

> More precisely, that it didn't compute the non-linear model. (the linear
> one is always computed since it's easy).  Perhaps check how that happens
> for tests/perfmodels/non_linear_regression_based.c vs how that doesn't
> seem to happen with your application. Normally that happens during
> starpu_shutdown in _starpu_deinitialize_registered_performance_models,
> in dump_reg_model that calls _starpu_regression_non_linear_power.

I experimented with non-linear models and my code indeed calibrates the 
non-linear models for some tasks. I am assuming that the GEMM tasks are 
simply too noisy. Is this plausible? I tried to calibrate the non-linear 
models on my desktop machine without GPUs and the GEMM tasks calibration 
still failed.

I will try to figure out why the GEMM tasks are so noisy. I think 
matrix-matrix multiplication should behave quite predictability.

 - Mirko

--
Mirko Myllykoski, Ph.D.
Senior research engineer, Department of Computing Science
Application expert, HPC2N
Umeå University

By sending an email to Umeå University, the University processes your 
personal data. For more information, please read: https://www.umu.se/en/gdpr